As Allison said, if you want performance you need to use the newer dispatch methods coupled with templating etc. The abstractions take a little effort to get your head around, but the performance gains are significant.
There is also the matter of algorithm design which is generally more important then tuning data access. In the old days most everything was serial. Again, rethinking basic approaches and redesigning parallel algorithms can have huge benefits. For example, I have routinely seen 10-100x gains (even a 1000x in one case) using new algorithm designs and the great dispatch methods that Allison introduced. The good news is that this newer coding style is making its way into VTK, so there are more examples to look at.
To help improve performance please specifically identify the filters that are considered too slow. Then the community can prioritize a list and we can work through the list to improve speeds.